Shieldstral is Mistral AI's 3B open-weight safety classifier for developers and trust-and-safety teams who need policy-adaptive moderation of text and images. Mistral released Shieldstral 1.0 on 4 August 2026 under Apache 2.0, and the design trick is that your policy is an input at inference rather than a taxonomy baked into the weights.
Core Features
- Moderation framed as binary question answering: you supply an instruction, one yes/no question and the content to judge.
- Returns a calibrated probability from the yes and no logits in one forward pass, so you pick the threshold.
- Native Pixtral vision encoder handles text, images, or text plus image in a single call.
- Built on Ministral-3-3B-Base and trained on 54.1M contrastive pairs across 12 languages.
- Runs in BF16 on one 16 GB GPU, served by vLLM, llama.cpp, SGLang or Transformers.
- Policies live in plain text, so changing a rule means editing a file instead of retraining.
Best For
- Regulated or data-residency-bound teams that cannot route user content through a third-party moderation API.
- Products needing different strictness per surface, strict on public comments and looser on internal drafting.
- Teams that want their harm taxonomy versioned in git next to application code.
Pricing
Weights are free at $0 under Apache 2.0 on Hugging Face, with no revenue cap and no hosted Shieldstral endpoint on La Plateforme at launch. Anyone quoting a per-call price is quoting their own hosting, so budget for a mid-range 16 GB GPU rather than an API bill.
Our Take
Mistral claims parity with guard models up to 7x its size, reporting 84.9% average text safety and 83.8% multimodal against Meta's Llama Guard and Google's ShieldGemma. Treat those as vendor-reported until third parties replay them. Best for self-hosted moderation you want to inspect and tune per surface. The trade-off is uneven non-English coverage, with Mistral flagging weaker Arabic and Indonesian classification. More options in AI models and engines on aifreetool.site.
